Shadow Removal

According to the algorithm of shadow removal, the paper [5] divides the method to two major architectures. First architecture is based on the result of background modeling and the threshold of parameter. In the method proposed by Cucchiara et al. [6], it divide the result of object detection to fore components, moving object, moving object shadow, ghost and ghost shadow in Fig. 1. Using the optical flow to eliminate the ghost and ghost shadow. This method considers the region of moving object shadow has the darker brightness and similar chromaticity. Using this property, the region with lower brightness and similar chromaticity in the foreground will define to the moving object shadow. In order to get more accurate moving object, this method has to remove the moving object shadow. The performance of remove the shadow of moving object is based on the threshold of the brightness. This method of shadow removal is refer the chromaticity and brightness. The other method use the property of environment

Fig. 1 The object classification result of [6]

proposed by Huang and Chen [7]. The light source can divide to two component, direct light and ambient light. Direct light is white like sun light and ambient light is more bluer than direct light. The cause of shadow is the direct light was blocked but still has ambient light. If the direct light was totally blocked, the region of shadow only has ambient light. So this method considers the change of color in the shadow region will between the full direct light and without direct light. In the training step, this method uses Gaussian model to build the shadow model base on the ground truth that label the background, foreground and shadow in whole scene. In addition to build the model with using color information, this method also build the model with the gradient in each pixel.

In the detect step. This method uses the weak detector to detect the foreground without the impossible shadow sample (e.g. the region with brighter color than background).

Using these Gaussian models to compute the probability of these pixel belonging to background, foreground or shadow. If the probability of these pixel belonging to foreground is larger than background and shadow, these pixel will be classify to foreground. This method uses the property of light source to detect the shadow by build the model to learn the change of light. Some method using this property will compute the relation between blue and other color information. The other method uses the information of texture was proposed by Sanin et al. [8]. This method considers the region of shadow has similar texture with background. Base on the result of object detection, this method use the chromaticity to choice the region is likely to shadow.

After select the candidate region of shadow, this method computes the gradient direction distance of the region pixel which has significant gradient magnitude. If the average gradient direction distance of region is lower than the threshold, the region will classified to the shadow. The performance of this method is based on the object detection and threshold of gradient direction distance. The method uses texture is

consider that has better performance than the method use color information. This architecture always attempt to adjust the parameter and threshold in the feature or find the different combined ratio of each feature to get the best performance in general scene.

There have a problem occur when the scene is very extreme. The performance will become poor than our expectation. In the previous paragraph, even the general scene may have great difference between each other. For example when the combination is suitable for the simple foreground scene, it may not suitable for the case in outdoor. It is hard to find the combination has best performance in every scene.

In order to find the good combination of each scene, second architecture proposed the method by using the classifier to build the model for each scene. The method proposed by Wang et al. [9] uses this architecture. This method uses LPGMM to detect the moving object first. LPGMM is the method of background modeling different from GMM with refer the information not only current pixel but also the neighbor pixel. This foreground includes some region of shadow was misclassified to foreground. So this method uses the SVM to learn the feature of shadow that can classify the foreground and shadow. Combining the object detection and classifier can get the more accurate foreground without the shadow. In this architecture, we need the ground truth data of the scene in order to label the feature of each pixel with it’s category (foreground, shadow or background) for the training stage. Because of these feature is label by the ground truth of each scene, we can build the model was suitable for each scene to classifier foreground and shadow.

In the propose method, we choose the feature those are suitable for different scene and use classifier to build the model in order to find the best performance of each scene.

Because we chose the feature are suitable for different scene, there has some feature has good performance and other has poor result in the same time. Base on the architecture

which using the classifier, the feature which has greater performance will get higher weight in the model by training use the classifier. So we can build the great shadow removal model for each scene.